Tensor-Network Inference in a Field-Coupled XY Model for Portfolio Allocation
Listen
Radio episode about this paper
Transcript
Introduction to the show: ident: Condensed Matter. The day in condensed matter: materials, phases and the measurements under them.
Kai: We're looking at the paper: "Tensor-Network Inference in a Field-Coupled XY Model for Portfolio Allocation".
Mira: Here's the gist of the paper. Tensor-network methods are applied to a field-coupled XY model for long-only portfolio construction We apply tensor-network methods to a field-coupled XY model for long-only portfolio construction This paper demonstrates a tractable route from financial time series to network-adjusted allocations through the use of tensor networks and statistical mechanics Model Formulation The core of the approach involves mapping portfolio construction onto an interacting XY model where asset angles are latent variables The field-coupled XY Hamiltonian is defined as H(θ) = −X(i,j)∈E Jij cos(θi − θj) − X i∈V hi cos θi This Hamiltonian incorporates empirical dependence through pairwise couplings Jij and local return signals via fields hi The Gibbs distribution p(θ) is then defined based on this Hamiltonian at inverse temperature β > 0, where the partition function Z is given by Z = Y k∈V Z 2π 0 dθ k 2π exp
−βH(θ): Interaction Graph Construction The sparse interaction path is constructed using a combination of correlation distance, four-point Gromov hyperbolicity, and Ward clustering Correlation distance is defined as dij = 2(1 − ρ bij) The four-point Gromov test compares the two largest of the three sums of opposite-pair correlation distances to diagnose tree-likeness Ward hierarchical clustering is used to obtain a concrete hierarchy whose leaf order defines the sparse interaction path, where consecutive leaves become neighbours on the interaction path Tensor Network Inference The continuous angular partition function is mapped to a finite-current tensor network using the Fourier–Bessel expansion This expansion replaces continuous angular integrals with integer-current representations, and the exact flux representation is derived in Equation (28) The numerical evaluation restricts every edge current to NK = (−K,..., K), resulting in a finite tensor-network contraction Z K Bottom-up and top-down contractions yield the one-site marginals mi = ⟨cos θi⟩, which are then converted into feasible capital weights wi = exp(γmi) P j exp(γmj) Results and Interpretation The study compares selected parameter pairs across five equity markets over continuous ranges of inverse temperature β and concentration γ The resulting portfolio weights w satisfy wi ≥ 0 and P i wi = 1, meaning the final allocation is long-only and fully invested The effective breadth Neff = 1/HHI is used as a diagnostic tool to characterize the portfolio concentration across different parameter regimes Several displayed XY portfolios exhibit favorable same-sample Sharpe ratios relative to benchmarks in certain markets, such as India where the (1, 1) portfolio achieves a Sharpe ratio of 1.240 The analysis confirms that while the model spans nearly equal-weight to strongly concentrated allocations, non-monotonic dependence on β at large γ can compress cross-sectional score differences Limitations and Scope The principal limitations of this construction include current truncation, replacement of the dense graph by a Ward-ordered path, same-sample parameter inspection, omitted trading costs, and the absence of rolling out-of-sample tests Predictive performance is explicitly outside their scope The comparison with Markowitz frontiers shows that the XY score map has a different objective than mean–variance optimization The construction's computational feasibility is demonstrated by the O(ND 2) scaling on the sparsified graph path, where D = 2K + 1 The study establishes the behaviour and computational feasibility of this construction rather than predicting future investment performance The paper successfully formulates long-only portfolio construction as an interacting XY model solved by tensor-network contraction Expected-return information enters through local fields, empirical dependence through pairwise couplings, and the equilibrium spin projections provide asset scores The final softmax maps these scores to positive weights that sum to one The XY angles are latent variables; capital allocation occurs only at this last step Several displayed XY portfolios have favourable same-sample Sharpe ratios relative to equal weight, the market index, minimum variance, and equal-risk contribution in certain markets The Markowitz tangency portfolio remains strongest because it maximises the same criterion on the evaluation sample The comparison shows the computational feasibility and parameter dependence of this construction This paper establishes the behaviour and computational feasibility of this construction rather than predicting future investment performance Acknowledgements Kartikeya Chowdhry acknowledges the use of Claude Opus 4.8 for assistance in writing code used in this study The authors declare that they have no known competing financial interests or personal relationships that could have appeared to influence the work reported in this paper A Supplementary Implementation and Universe Details section provides exact column lists for the complete-case matrices used for estimation across all five markets The empirical study uses fixed large-cap universes from the United States, United Kingdom, Germany, Japan, and India The stored snapshot was downloaded through yfinance; securities satisfy the 97% availability rule, and the complete retained lists are reported in Appendix A Benchmarks use daily simple returns and annualisation by 252; log returns supply the XY correlations and fields The inference and allocation maps are defined for continuous parameter values β > 0 and γ ≥ 0 The concentration surfaces in Section 6.4 sample β ∈
0.2, 16: and γ ∈
0, 120: ; the contraction and softmax procedures are unchanged at every point in this range The six representative pairs cover near-equal, intermediate, and concentrated allocations These points sample a continuum and were chosen to show qualitatively different regimes, from nearly uniform to strongly concentrated allocations The highest-Sharpe displayed XY points combine annualised return and volatility of 27.0% and 20.9% in the US, 21.8% and 27.3% in the UK, 15.0% and 18.0% in Germany, 27.1% and 24.0% in Japan, and 19.9% and 16.
Title and authors: Kai: So, we're looking at this paper called "Tensor-Network Inference in a Field-Coupled XY Model for Portfolio Allocation." It sounds a bit dense, but basically, they’re taking financial data—like daily returns—and trying to figure out how to build a portfolio from that.
Mira: Yeah, it’s about mapping asset angles onto this interacting system where the angles are latent variables. It suggests you can get network-adjusted allocations directly from time series data using these tensor network methods and statistical mechanics.
Lev: From my side, the idea is mapping continuous allocation problems onto a discrete structure that we can actually solve with current hardware constraints. It’s about finding a tractable route from raw financial data to actual portfolio weights without needing some kind of complex optimization step at the end.
Kai: Exactly. The core idea is using this field-coupled XY Hamiltonian where you have these empirical dependence terms and local return signals acting as fields on the assets, and then you solve that system through a tensor network contraction.
Mira: That's the mechanism they use to model how assets influence each other through their correlations while also incorporating those specific return signals for each asset.
Lev: And the way they construct this interaction graph using correlation distance and Ward clustering is key because it makes the problem sparse, which is what you need for any kind of computation, especially on real hardware.
The paper's summary: Kai: So, looking at the summary of "Tensor-Network Inference in a Field-Coupled XY Model for Portfolio Allocation," they're showing how they use this framework to handle long-only portfolio construction by treating asset angles as latent variables.
Mira: They explain that you define the field coupled XY Hamiltonian, and then the Gibbs distribution is set up using an inverse temperature beta greater than zero. The partition function Z is defined over these angular variables, and they show how this maps to a finite-current tensor network using a Fourier–Bessel expansion.
Lev: The crucial step there is that they restrict every edge current to some maximum value K, which means every bond in the tensor network has a finite dimension, making the numerical evaluation possible. This gives them those one-site marginals, or the equilibrium cosine scores, through bottom-up and top-down contractions.
Kai: Those scores are then converted into feasible capital weights using a softmax map that involves an inverse temperature gamma and those marginals. That's how they get the final allocation from the model’s equilibrium state.
Mira: And they test this across five equity markets over continuous ranges of beta and gamma, showing that their resulting portfolio weights satisfy two things: they are positive, meaning long-only, and they sum to one, meaning fully invested.
Lev: The paper claims this provides a tractable route from financial time series to network adjusted allocations by using these methods, even though it explicitly states that predictive trading performance is outside the scope of what this construction can do.
The paper's improvements: Kai: They point out several key improvements in their approach. First, they refine the construction of the sparse interaction path by combining correlation distance, four-point Gromov hyperbolicity, and Ward clustering to define that specific network structure.
Mira: That’s designed to ensure that when you build this interaction graph, you get a path where consecutive leaves are neighbors on that path, which keeps the interactions relatively strong. It’s a way to make the sparse approximation more meaningful.
Lev: From an engineering standpoint, they use the Fourier–Bessel expansion to map the continuous angular partition function into integer-current representations, which is essential because it allows for a finite tensor network contraction instead of dealing with an infinite sum.
Kai: And then they show that this bottom-up and top-down contraction process successfully yields those one-site marginals, or the equilibrium cosine scores, which are the key inputs for the final allocation.
Mira: They also emphasize how restricting the angles to two opposite orientations recovers a discrete Ising state space, which is useful because it links their continuous model back to formulations that are often used for discrete financial decisions like asset inclusion.
Lev: The finite-current cutoff they use, where they restrict the absolute value of the current to less than or equal to K, is what makes the entire contraction numerically possible and scalable on a computer.
Conclusion: Kai: So, wrapping up the paper "Tensor-Network Inference in a Field-Coupled XY Model for Portfolio Allocation," they show this construction works by using empirical dependence and local return signals to generate scores that map to positive, fully invested weights.
Mira: The main implication they’re pushing is that this provides a tractable route from financial time series directly to network adjusted allocations through statistical mechanics, even though it's not designed for predicting future trading performance.
Lev: For someone building real hardware, what matters is the computational feasibility they demonstrate; the scaling on the sparsified graph path is shown to be O of n squared where n is related to K, which suggests it’s manageable.
Kai: And they show that even though you can go from nearly equal-weight to strongly concentrated allocations by changing parameters like beta and gamma, there are still favorable same-sample Sharpe ratios in certain markets for the XY portfolios.
Mira: They note that the model confirms this behavior, showing how non-monotonic dependence on beta at large gamma can actually compress those cross-sectional score differences between different assets.
Lev: From an error correction view, they show that while this is a model of equilibrium, it’s not a projection or constrained optimization step for actual trading, which keeps the overhead minimal for running on hardware.
Kai: So that's it for this paper on "Tensor-Network Inference in a Field-Coupled XY Model for Portfolio Allocation." It establishes how to use these techniques to generate portfolio weights from financial data by solving the underlying interacting XY model.
Kartikeya Chowdhry, Dr. G Subrahmanya V.R.K. Rao
Gopal K. Saxena Centre for Advanced Research
cond-mat.stat-mech
Submitted: 2026-09-04
Updated: 2026-10-30
Comments: 52 pages, 10 figures, 5 tables
License: http://arxiv.org/licenses/nonexclusive-distrib/1.0/
Importance score: 75/100
The gist: Tensor-network methods are applied to a field-coupled XY model for long-only portfolio construction We apply tensor-network methods to a field-coupled XY model for long-only portfolio construction…
Key concepts
- Field-Coupled XY Hamiltonian
- This mathematical model describes the system where asset angles are the 'spins.' It includes terms for pairwise dependencies (couplings) between assets and local signals (fields) representing individual asset returns. This allows the model to capture how assets influence each other's orientations based on empirical data.
- Tensor Network Inference
- This technique is used to solve the complex continuous angular partition function by mapping it onto a finite tensor network. This allows for numerical evaluation of asset scores (marginal probabilities) through specific contractions, turning a continuous problem into a solvable, discrete calculation.
- Softmax Mapping for Allocation
- The final step uses a softmax function to convert the calculated equilibrium scores (derived from the model) into actual capital weights. This mapping ensures that all resulting weights are positive and sum exactly to one, guaranteeing the portfolio is long-only and fully invested.
Terminology
Summary
Tensor-network methods are applied to a field-coupled XY model for long-only portfolio construction We apply tensor-network methods to a field-coupled XY model for long-only portfolio construction This paper demonstrates a tractable route from financial time series to network-adjusted allocations through the use of tensor networks and statistical mechanics
Model Formulation
The core of the approach involves mapping portfolio construction onto an interacting XY model where asset angles are latent variables The field-coupled XY Hamiltonian is defined as H(θ) = −X(i,j)∈E Jij cos(θi − θj) − X i∈V hi cos θi This Hamiltonian incorporates empirical dependence through pairwise couplings Jij and local return signals via fields hi The Gibbs distribution p(θ) is then defined based on this Hamiltonian at inverse temperature β > 0, where the partition function Z is given by Z = Y k∈V Z 2π 0 dθ k 2π exp[−βH(θ)]
Interaction Graph Construction
The sparse interaction path is constructed using a combination of correlation distance, four-point Gromov hyperbolicity, and Ward clustering Correlation distance is defined as dij = 2(1 − ρ bij) The four-point Gromov test compares the two largest of the three sums of opposite-pair correlation distances to diagnose tree-likeness Ward hierarchical clustering is used to obtain a concrete hierarchy whose leaf order defines the sparse interaction path, where consecutive leaves become neighbours on the interaction path
Tensor Network Inference
The continuous angular partition function is mapped to a finite-current tensor network using the Fourier–Bessel expansion This expansion replaces continuous angular integrals with integer-current representations, and the exact flux representation is derived in Equation (28) The numerical evaluation restricts every edge current to NK = (−K,..., K), resulting in a finite tensor-network contraction Z K Bottom-up and top-down contractions yield the one-site marginals mi = ⟨cos θi⟩, which are then converted into feasible capital weights wi = exp(γmi) P j exp(γmj)
Results and Interpretation
The study compares selected parameter pairs across five equity markets over continuous ranges of inverse temperature β and concentration γ The resulting portfolio weights w satisfy wi ≥ 0 and P i wi = 1, meaning the final allocation is long-only and fully invested The effective breadth Neff = 1/HHI is used as a diagnostic tool to characterize the portfolio concentration across different parameter regimes Several displayed XY portfolios exhibit favorable same-sample Sharpe ratios relative to benchmarks in certain markets, such as India where the (1, 1) portfolio achieves a Sharpe ratio of 1.240 The analysis confirms that while the model spans nearly equal-weight to strongly concentrated allocations, non-monotonic dependence on β at large γ can compress cross-sectional score differences
Limitations and Scope
The principal limitations of this construction include current truncation, replacement of the dense graph by a Ward-ordered path, same-sample parameter inspection, omitted trading costs, and the absence of rolling out-of-sample tests Predictive performance is explicitly outside their scope The comparison with Markowitz frontiers shows that the XY score map has a different objective than mean–variance optimization The construction's computational feasibility is demonstrated by the O(ND 2) scaling on the sparsified graph path, where D = 2K + 1 The study establishes the behaviour and computational feasibility of this construction rather than predicting future investment performance
Conclusion
The paper successfully formulates long-only portfolio construction as an interacting XY model solved by tensor-network contraction Expected-return information enters through local fields, empirical dependence through pairwise couplings, and the equilibrium spin projections provide asset scores The final softmax maps these scores to positive weights that sum to one The XY angles are latent variables; capital allocation occurs only at this last step Several displayed XY portfolios have favourable same-sample Sharpe ratios relative to equal weight, the market index, minimum variance, and equal-risk contribution in certain markets The Markowitz tangency portfolio remains strongest because it maximises the same criterion on the evaluation sample The comparison shows the computational feasibility and parameter dependence of this construction This paper establishes the behaviour and computational feasibility of this construction rather than predicting future investment performance
Acknowledgements
Kartikeya Chowdhry acknowledges the use of Claude Opus 4.8 for assistance in writing code used in this study The authors declare that they have no known competing financial interests or personal relationships that could have appeared to influence the work reported in this paper A Supplementary Implementation and Universe Details section provides exact column lists for the complete-case matrices used for estimation across all five markets The empirical study uses fixed large-cap universes from the United States, United Kingdom, Germany, Japan, and India The stored snapshot was downloaded through yfinance; securities satisfy the 97% availability rule, and the complete retained lists are reported in Appendix A Benchmarks use daily simple returns and annualisation by 252; log returns supply the XY correlations and fields The inference and allocation maps are defined for continuous parameter values β > 0 and γ ≥ 0 The concentration surfaces in Section 6.4 sample β ∈ [0. p. 02, 16] and γ ∈ [0, 120]; the contraction and softmax procedures are unchanged at every point in this range The six representative pairs cover near-equal, intermediate, and concentrated allocations These points sample a continuum and were chosen to show qualitatively different regimes, from nearly uniform to strongly concentrated allocations The highest-Sharpe displayed XY points combine annualised return and volatility of 27.0% and 20.9% in the US, 21.8% and 27.3% in the UK, 15.0% and 18.0% in Germany, 27.1% and 24.0% in Japan, and 19.9% and 16.
Green — the paper says it, and the page is printed. Red — no passage carries it, so treat that sentence as unverified. The same check runs on your own text: check a summary against its source, $5.
Improvements for AI systems
-
Correlation network construction for sparse interaction paths improves asset selection by using Ward hierarchical clustering to define an interaction path based on correlation distance, which
defines a sparse interaction path
and keepsrelatively strong consecutive-leaf pairs.
-
Continuous allocation approximation via Fourier-Bessel expansion enables the mapping of the continuous angular partition function to a finite-current tensor network, allowing for numerical evaluation of marginals and scores through bottom-up and top-down contractions.
-
The use of the field coupled XY Hamiltonian allows AI systems to model asset interactions where
positive correlation couplings promote alignment of neighbouring latent states
and local fields encodeasset-specific signals,
leading to network adjusted scores that are then converted intopositive, fully invested portfolio weights.
-
The softmax map converts equilibrium scores into feasible capital weights, ensuring the resulting allocations are long-only and fully invested by using the formula wi = exp(gamma mi) / sum j exp(gamma mj), which guarantees "wi > 0
and that the portfolio is
long-only and fully invested without a projection or constrained optimisation step." -
The model provides a tractable route from financial time series to network adjusted allocations, allowing AI systems to generate portfolio weights by applying the workflow:
The correlation-distance geometry determines the Ward hierarchy and interaction path; the XY angles belong to the separate spin space on which Gibbs inference is performed.
Sources
Related papers
- Emergent long-tail dynamics in driven magnets with dynamical frustration
- Statistical Mechanics of Dynamical System Identification
- Second law of thermodynamics in closed quantum many-body systems
- Chemical potentials from structure factors: II. Charged multi-component mixtures
- Genuine quantum scars in Floquet chaotic many-body systems
- Stochastic Thermodynamics for Autoregressive Generative Models: A Non-Markovian Perspective